Related Experiment Video
Updated: Aug 14, 2025
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
48.2K
Automatic detection and vascular territory classification of hyperacute staged ischemic stroke on diffusion weighted
Kun-Yu Lee1,2, Chia-Chuan Liu1,2, David Yen-Ting Chen1,2
1Department of Medical Image, Shuang Ho Hospital, Taipei Medical University, No. 291, Zhongzheng Road, Zhonghe District, New Taipei City, 23561, Taiwan, ROC.
Scientific Reports
|January 9, 2023
Summary
Convolutional neural network (CNN) models accurately detect and classify ischemic stroke in early stages using diffusion-weighted images (DWI). Inception-v3 demonstrated the highest performance, aiding in timely mechanical thrombectomy decisions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Automated detection and classification of ischemic stroke in hyperacute stages are crucial for effective mechanical thrombectomy.
- Diffusion-weighted imaging (DWI) is essential for evaluating stroke, particularly in the hyperacute phase.
Purpose of the Study:
- To evaluate the performance of various convolutional neural network (CNN) models for detecting ischemic stroke.
- To classify hyperacute ischemic stroke into anterior circulation infarct (ACI), posterior circulation infarct (PCI), or normal image slices using DWI.
Main Methods:
- A retrospective study using 253 hyperacute DWI cases, with 127 cases (2119 slices) ultimately included.
- Three CNN models (Inception-v3, EfficientNet-b0, modified LeNet) were trained and evaluated on ACI, PCI, and normal DWI slices.
- Gradient-weighted class activation mapping (Grad-Cam) was used to generate activation maps for model interpretability.
Main Results:
- Inception-v3 achieved the highest accuracy (86.3%), weighted F1 score (86.2%), and kappa score (0.715).
- The modified LeNet model showed strong performance with 85.2% accuracy and 84.7% weighted F1 score.
- EfficientNet-b0 had the lowest performance among the tested models (83.6% accuracy).
- Activation maps indicated susceptibility artifacts as a potential cause for misclassification.
Conclusions:
- CNN models can achieve high performance in detecting and classifying ischemic stroke on hyperacute DWI according to vascular territory.
- Inception-v3 shows significant promise for automated stroke detection and classification in clinical settings.
- Further research may focus on mitigating misclassifications caused by imaging artifacts.
